Tuesday, June 20, 2017

Machine learning basics

There are several types of machine learning but we will focus on the followings in this article:

  • Supervised learning
  • Unsupervised learning
  • Reinforcement learning

Supervised learning

Basically used to make predictions about future data from labeled/categorized training dataset.

A dataset is a table where:

  • each row is a sample
  • each column is a feature
  • each row is labeled with a class label


A supervised learning task with discrete class labels is called a classification task.
Classification is a subcategory of supervised learning where the goal is to predict the categorical class labels of new instances based on past observations.


Dataset with  [data; label]
Data 0 : [ I like sport; Present]
Data 1 : [ I love shopping, Present]
Data 2 : [ I was in Amsterdam, Past]
Data 3 : [ I did something wrong, Past]
Data 4 : [ I do exercise every day, Present]

New data:

I did not know --> Past
I love running --> Present

With 2 possible class labels, the task is a binary classification task.
With more, it is a mutli-class classification task.

Example of multi-class dataset:

[Picture of cat; cat]
[Picture of dog; dog]
[Picture of mouse; mouse]

Here the machine learning system would be able to recognize a dog, a cat or a mouse but wouldn't succeed with any other animal because it is not part of our dataset.

Typical example of two-dimensionnal dataset for a binary classification task:

Data 0 : [  [0;0] ; Orange]
Data 1 : [  [1;1.5] ; Orange]
Data 2 : [  [1;2] ; Orange]
Data 3 : [  [1;2.8] ; Orange]
Data 4 : [  [2;1.5] ; Orange]
Data 5 : [  [2;2.5] ; Orange]
Data 6 : [  [3;0] ; Blue]
Data 7 : [  [3;1.5] ;Blue]
Data 8 : [  [3;2] ; Blue]
Data 9 : [  [4;2.8] ; Blue]
Data 10 : [  [4;1.5] ; Blue]
Data 11 : [  [4;2.5] ; Blue]
Data 12 : [  [5;3] ; Blue]

It is two-dimensionnal because each sample of the dataset has 2 values (usually named x1,x2). If we represent these samples on a 2-dimensionnal graph, we would see this:

The prediction would be based on the distribution of the sample. A point with x1 > 3 would be predicted as Blue and a point with x1 < 2 would potentially be red.


Regression is also called prediction of continuous outcomes. In regression analysis we give a serie of numbers (x or predictor) and response variables (y or outcome) and we try to find a relationship between them to predict a future outcome.


with [x;y]
Data 0 : [ 0 ; 0 ]
Data 1 : [ 1 ; 1.5 ]
Data 2 : [ 1.5 ; 1 ]
Data 3 : [ 2 ; 2 ]
Data 4 : [ 2.5 ; 2.6 ]
Data 5 : [ 3 ; 3.2 ]
Data 5 : [ 4 ; 3.9 ]

Several types of algorithm can be selected to process input data. The following figure illustrates the concept of linear regression:

The computed curve will be used to predict the outcome of new data.

Reinforcement learning

Here the goal is to develop a system (agent) that improves its performances based on interactions with environment. The system will receive a feedback (reward) for every one of its actions. Each reward informs him of the quality of his action.
The agent will learn a series of actions that maximizes this reward via an empirical try-and-error approach.

A typical example is Google's Deepmind which beat the best Go players.

Unsupervised learning

In supervised learning, we include the right answer (labels) into the dataset. Here, we don't know the right answer beforehand. We are dealing with uncategorized data with an unknown structure.
With unsupervised learning, we can explore the structure of our data to extract meaningful information without an outcome or a reward.


Clustering is an exploratory data analysis technique which groups data together by similarity (unnsupervised classification).

Dimensionality reduction

Dimensionality reduction is another unsupervised learning field. To prevent against the computation of huge amounts of data which results in performance and storage issues, unsupervised dimensionality reduction preprocesses data to remove noise and retain relevant information.

Read More

Thursday, June 15, 2017

Django + Js : display counting timedelta in view

Suppose you have an entry in DB with a timestamp 'start_date'. You want to display the time elapsed since 'start_date' in your view and you want that delay to grow in real time like a clock. In the following example, the user can select one entry at a time and the counter has to be updated accordingly.

Note: To achieve this it is important to work with UTC timestamps in both DB and your view

Here is what my code in django views.py looks like:

def get_entry_info(request):
    Recover information for the specified entry name
    entry_name = request.POST.get("name", None)
    if request.method == "POST" and entry_name is not None:
        data = {}
        # Get entry object 
        entry = MyEntries.objects.get(name=entry_name)
        data['start_date'] = int(time.mktime(entry.start_date.replace(tzinfo=None).timetuple())) * 1000 
        return JsonResponse(data)
        return HttpResponse("Invalid entry name")

Here is what my javascript code looks like:
var startDate;

// Loads content into the information panel
// @param data : data to be displayed in the information panel
function reloadInformationPanel(entryname){
    // Get job information from server
        headers: { "X-CSRFToken": '{{ csrf_token }}' },
        url: "{% url 'get_entry_info' %}",
        method: 'POST', 
        dataType: 'json',
        data: {
            'name': entryname, // outgoing data
        success: function (data) {        
            startDate = new Date(data.start_date);
        error: function(xhr,errmsg,err) {
            console.log(xhr.status + ": " + xhr.responseText); 

// Starts timer of job duration
function startTime() {
   var now = convertDateToUTC(new Date());
   var delay = new Date(now - startDate);
   document.getElementById('entry_duration').innerHTML = delay.getUTCHours() + "h" + delay.getUTCMinutes() + "m" + delay.getUTCSeconds() + "s";
   var t = setTimeout(startTime, 500);

// Converts a Date object to UTC
function convertDateToUTC(date) { 
   return new Date(date.getUTCFullYear(), date.getUTCMonth(), date.getUTCDate(), date.getUTCHours(), date.getUTCMinutes(), date.getUTCSeconds()); 

Now simply map reloadInformationPanel to the entry selection button ;)

Read More